When an existing class's catch-all bucket accumulates sufficient unclassified claims, the Custodian analyzes them with an LLM to identify shared mechanisms and generates candidate factors that remain in a pending-approval state until human review.
Referenced by
